On May 19, 2025, Jensen Huang stood on a stage in Taipei and said NVIDIA is no longer a chip company — it builds "AI infrastructure." For an Indian AI startup paying every month for a single rented A100, that framing decides your 2026 burn rate. This post pulls the four announcements that actually change your inference bill and the one decision they force: self-host or keep renting.
The 60-Word Answer
NVIDIA's Computex 2025 keynote (May 19) confirmed GB300 for Q3 2025 with 1.5x more inference, 1.5x more HBM, and 2x more networking than GB200. It opened NVLink to rival chips through NVLink Fusion, and pushed two desktop AI boxes — DGX Spark and DGX Station. For most Indian startups under 50 daily inference users, the practical takeaway is unchanged: rent first, self-host only when your monthly GPU bill is large and steady.
Why This Matters Now (May 2025)
Computex ran May 19–21 in Taipei, and Huang's keynote opened it. The reason it lands for Indian founders: GPU rental prices in India trail the global curve by 6–9 months, so a roadmap announcement today is a price signal for your Q4 budget. Huang's "AI factory" pitch — covered by CNBC's keynote report — is a sell for buying racks, not single cards. Read it through that lens.
What NVIDIA Actually Announced
Four things matter for a startup buying or renting compute. We've stripped out the data-center theatre.
The single most interesting line, per DigiTimes' NVLink Fusion coverage, is that the ASIC surge pushed NVIDIA to open NVLink at all. Custom inference chips were eating into NVIDIA's moat. Opening the interconnect is a defensive move — and it's good news if you ever want a non-NVIDIA accelerator in your stack without losing the software ecosystem.
The GPU-Cost Math for an Indian Startup
Here's the comparison that actually decides things. We ran inference for an in-house product and three client builds through 2025, so these are real Indian-market costs, not US list prices.
Read the chart carefully. On-demand A100 rental in India is the most expensive option for a steady workload — you pay for idle hours. The API route wins at low and bursty volume because someone else eats the idle cost. Owning hardware only pays back when your utilisation crosses roughly 60% of a 24-month amortisation window. The GB300 news does not change this logic; it changes the ceiling of what a single owned box can do in 2026.
When to Self-Host vs Rent Inference (DIY Decision)
This is the walkthrough. Run it in an afternoon with your last three months of token logs.
We applied this exact framework to our voice-AI work. On TalkDrill, our in-house English-speaking app, the speech models run hot enough during Indian evening peak that owning part of the stack made sense, while the long-tail off-peak hours stayed on metered APIs. The split — own the floor, rent the spikes — is the pattern we now recommend most.
What NVLink Fusion Quietly Changes for Your 2026 Stack
The interconnect news reads like data-center plumbing, but it has a real effect on costs two years out. As Hrishikesh, our CTO, frames it: NVIDIA opened NVLink because custom inference silicon was getting good enough to threaten its margin. That competition is what eventually lowers prices for buyers — but only after the software stack catches up.
Here's the practical read for an Indian startup planning hardware in 2026. Today, the NVIDIA software ecosystem — CUDA, the driver stack, the optimised inference libraries — is the real lock-in, not the chip. NVLink Fusion doesn't change that overnight. What it does is create a credible path where a future accelerator from Marvell, Alchip, or a hyperscaler's own ASIC can slot into an NVLink fabric without you rewriting your whole serving layer.
If you're a small team, treat this as a watch item with one trigger: when a non-NVIDIA accelerator ships with a mature, documented inference runtime that your framework supports out of the box, re-price your options. Until then, the cheapest token for most Indian workloads still comes from a well-utilised NVIDIA card or a metered API — the same conclusion the cost math gave us above. The roadmap widens your future menu; it doesn't change today's order.
Your Buy-vs-Rent Checklist
Run this before you sign any GPU purchase order after a NVIDIA keynote.
- You have 90 days of real token-volume data, with p50 and p95 daily figures
- You priced the API path at peak (p95) volume, not median
- You priced owned hardware fully loaded: card ÷ 24 months + power + rack
- You computed actual utilisation (active inference hours ÷ 24) — and it clears 60%
- You budgeted 0.2 FTE of engineer time for driver patching and uptime
- You checked the used A100/H100 market timing against the new generation's ship date
- You have a written volume trigger that flips the decision (e.g. median > 6M tokens/day)
- You have a power-backup plan if you're putting an owned box in an office, not a colo
Common Mistakes (When the Roadmap Misleads You)
A fourth trap: assuming NVLink Fusion means you can mix vendors cheaply tomorrow. It opens the door, but the software, drivers, and support maturity for non-NVIDIA accelerators in India is still thin. Treat it as a 2027 option, not a 2025 plan.
A Real Example: Pricing Inference for a Pune Document-AI Startup
A 7-person Pune startup parses insurance documents with an open-weight model. Their question after Computex: should they buy a GPU box now? We ran the framework. Median volume was 1.1M tokens/day, p95 was 4.3M — very bursty, because claims arrive in waves after weekends and month-ends.
This is the same self-host-vs-cloud reasoning we use for automation infrastructure with our AI and automation team — the tool changes, the unit-economics question doesn't.
FAQ
Should an Indian AI startup buy GPUs after Computex 2025?
Usually not yet. GB300 ships Q3 2025, which softens the used A100/H100 market by early 2026. If your GPU utilisation is under 60% of a 24-month amortisation, renting or using an API is cheaper. Buy on a volume trigger, not on keynote excitement.
What was the biggest Computex 2025 announcement for builders?
NVLink Fusion. NVIDIA opened its high-speed interconnect to non-NVIDIA CPUs and custom ASICs, with partners including MediaTek, Marvell, Alchip, Astera Labs, Synopsys, and Cadence. It signals more mixed-silicon options ahead, though production maturity for non-NVIDIA chips in India lags by a year or two.
Is the DGX Spark good for an Indian startup?
For development and small inference, yes. It's a desktop AI box NVIDIA said is shipping within weeks of Computex. It is not a production high-availability setup — a single box is exposed to power cuts, which are frequent in much of India. Use it to prototype, not to serve customers.
How much does it cost to rent an A100 in India?
It depends on provider and commitment, but on-demand A100 80GB rental in India through 2025 was the most expensive option for a steady workload because you pay for idle hours. Spot and committed-use discounts cut it, but bursty workloads almost always do better on metered APIs.
Does GB300 change the self-host math?
Not the logic, only the ceiling. The utilisation break-even (~60%) stays the same. GB300 raises how much one owned box can do and, more usefully, pushes down used-market prices for the prior generation 6–9 months after it ships. That timing is the real lever for cost-conscious Indian teams.
What's the safest way to act on a chip roadmap?
Run a token-volume audit, price API vs owned at your peak, compute utilisation, and lock the cheaper option for two quarters. Re-decide each NVIDIA generation. The roadmap is a price signal, not a buy order — treat it like one.
Want an Honest Inference-Cost Review?
We run a 90-minute audit of your token volume and quality needs, then tell you whether to stay on an API, rent, or buy — with the rupee math, not a sales pitch. Suitable if you're an Indian startup with a growing inference bill. First call is with the engineer who'd own your stack.
Book an Inference-Cost ReviewRelated reading
